Two Teletype® Model 33 terminals, side by side, driven by one keyboard. One is wired to ELIZA (1966). One is wired to a local LLM (2026). Which is which is randomised, and you have to commit to a guess before the answer is shown.
Both machines were typed the same thing. One of those answers is sixty years older than the other. Which one?
The machine is not the product. The blind test is the product. The machine is the stage — and it happens to be the best teaching device ever built for explaining why terminals behave the way they do.
If you have ever wondered why \r and \n are two different characters, why
Ctrl+C interrupts, why the terminal beeps, or what stty erase is for: they
are all fossils of this one machine, and you can hear every one of them.
Everything, including the LLM, in one command:
docker compose up --buildthen open http://localhost:16001. The model is inside the image, so there is nothing to download at boot.
Where huggingface.co is unreachable:
docker compose build --build-arg HF_HOST=https://hf-mirror.comES modules do not load over file://, so opening web/index.html by
double-click will not work.
Start it once with the dev overlay:
docker compose -f compose.yml -f compose.dev.yml upThat bind-mounts web/ over the image's static directory. Edit a file,
refresh the browser — that is the whole loop. No rebuild, no restart, and
the LLM container is never touched.
The overlay is deliberately not named docker-compose.override.yml: that
filename is loaded automatically, and committing it would silently make the
production server serve files from its own host directory.
| Develop | docker compose -f compose.yml -f compose.dev.yml up |
| Front end alone, no containers | npm run serve |
| Rebuild the front-end image only | docker compose build web (seconds) |
| Production | docker compose up --build -d |
| Change the model | edit MODEL_FILE, then docker compose build llm |
| Tests | npm test |
The model is not re-downloaded when you change code. It lives in a
different image (Dockerfile.llm), and even rebuilding that one hits the
Docker layer cache unless MODEL_REPO, MODEL_FILE or HF_HOST changed.
Never mount code with a named volume. A named volume over a path the image already populates is seeded once, on first creation, and never updates — you would rebuild, restart, and still be served stale code with nothing to indicate why. Bind mounts do not have this problem.
Every constraint of the machine is load-bearing. Removing one to make the thing nicer removes the point.
| Real | |
|---|---|
| 72 characters per line | terminfo cols#72 |
| 10 characters per second, 110 baud | Model 33 spec |
| 6 lines per inch, 8.44 inch roll paper | Model 33 spec |
| Uppercase only, 64-character ASCII | there is no lowercase on the type drum |
| No backspace key | the carriage moves forward, or it returns |
RUB OUT is not backspace |
it is DEL, for punching out tape holes |
LINE FEED and RETURN are separate keys |
two mechanical actions — which is why CR and LF are two characters, and why Windows kept CRLF |
@ is SHIFT+P, not its own key |
Ray Tomlinson picked it off this keyboard in 1971 |
| No auto-wrap at column 72 | it overstrikes in place; wrapping is the host's job |
| The bell rings near end of line | warns the typist — and ASCII 0x07 rings it too |
| A carriage return takes real time | which is why real hosts padded with NULs after CR |
The terminfo entry for this machine is the one used as the worked example in
terminfo(5):
33|tty33|tty|model 33 teletype,
bel=^G, cols#72, cr=^M, cud1=^J, hc, ind=^J, os,
hc = hardcopy. os = overstrike. Both are implemented.
| Deliberately not real | Why |
|---|---|
| The cheat panel | so remote echo can be shown side by side on video. Labelled as such |
READY after every reply |
a timesharing BASIC prompt, not something ELIZA on CTSS printed. Kept because both hosts must format output identically |
| 30 and 120 CPS buttons | mercy for the impatient |
Some details are unverified and marked as such in the source rather than
asserted: several of the control-code keycap legends, which row HERE IS /
RUB OUT / REPT / BREAK belong to, and the keyboard row stagger.
| Type this | What happens |
|---|---|
PASSWORD ⏎ |
both hosts stop echoing. Keep typing — nothing appears on either roll. This is why your password does not echo in a terminal today |
HELLO###P ⏎ |
there is no backspace. # erases one character, @ kills the line, and both of them print. The paper keeps an honest record of every mistake |
HELP ⏎ |
the machine explains itself, on paper |
| Backspace | the window flinches and nothing happens |
PRINT |
the two rolls, one page each, at real geometry — 8.44in paper, 10 CPI, 6 LPI, torn top and bottom |
With headphones, A is in the left channel and B in the right. All audio is synthesised live in Web Audio: no samples, no files. The two motors are tuned 60 Hz and 60.4 Hz so they beat against each other, like two real machines on one bench.
web/ native ES modules, no bundler
├── index.html
├── css/ base, machine, keyboard, panel, dialog, print
└── js/
├── main.js boot and wiring only
├── config.js the numbers nothing else may invent
├── clock.js THE metronome -- one timer, both machines
├── machine.js roll, queue, strike, carriage
├── audio.js Web Audio synthesis
├── keyboard.js layout data + press animation
├── format.js THE shared post-processor <- read this one
├── hosts.js registry + sync gate
├── eliza.js DOCTOR engine, 1966
├── llm.js llama-server client, 2026
├── game.js assignment, guessing, score
├── print.js print sheet builder
├── eliza-rules.js the 1966 script, verbatim, + its reader
└── i18n.js + locales/{en,zh,ja}.js
Dockerfile nginx + web/
Dockerfile.llm llama-server + the model, baked in
nginx.conf static files, and /v1 -> llama-server
test/ node --test, no framework
docs/ model33.md, eliza.md
docs/eliza-original/ the 1966 script as published. Archive: do not edit
There is one deployment and one artefact. A build without the LLM would be a build without the blind test, and the blind test is the product.
npm test runs the suite. It covers exactly one module, format.js, because
that is the only part of this project where a silent bug is a product bug —
see below.
The whole thing collapses if the machine leaks which host is which, and latency is only the most obvious channel.
- Sync gate. Both hosts are resolved with
Promise.alland only then queued, so both start printing on the same frame. ELIZA answers in 0 ms and the LLM in 1–3 seconds; without this, response time alone gives it away instantly. (hosts.js) - Shared metronome. One timer steps both machines. They cannot run at
different speeds and cannot drift apart. (
clock.js) - One shared post-processor. Uppercase, filter to the 64-character set,
wrap at 72, one leading blank line, one trailing
READY. Neither host formats its own output. Anything about the shape of a reply is a tell independent of its content. (format.js) - One guess per round.
REVEALdisables itself after a guess; onlyRESHUFFLEre-arms it. - Tear-off does not reshuffle. New paper, same assignment. Tearing must not hand you a fresh bet.
The filter has to be code, not a system prompt. A small model will emit lowercase, Markdown, emoji, curly quotes and em dashes no matter what you ask it, and the machine physically cannot print any of that.
Two different things, so two different functions in format.js:
downgrade() |
hardware. Uppercase, fold curly quotes and dashes, filter to 0x20–0x5F. What the type drum carries |
houseStyle() |
convention. Markdown, punctuation vocabulary. What a HOST may print |
The distinction is load-bearing. The drum really does carry ? and *, and
the keyboard really can type them — echo does not pass through houseStyle.
It is only the two computers that may not use them.
Punctuation vocabulary. Measured across all 199 reassembly templates in the 1966 script, ELIZA's entire punctuation vocabulary is four characters:
' x34 , x13 - x5 . x1
No double quote. No question mark — not one, anywhere in the script. The
model reaches for all of them, so any mark it uses that ELIZA cannot is a
character that appears on exactly one roll. Host output is folded into that
vocabulary: " → ' (ELIZA quotes too, it just quotes the 1966 way — WHY DO YOU SAY 'AM'), ;: → ,, ?! → ., anything else dropped.
No terminal punctuation. 197 of the 199 templates end on a letter or a
digit and the other two on an apostrophe; none end on .?!. The model ends
almost every sentence with ?. That is a one-character tell — you could
win by looking at the last character of the roll without reading a word.
Word budget of 20. One more than the longest reply ELIZA was measured producing over a real conversation. The model's median was 18 words against ELIZA's 8. Trimming spends the budget back to the last clean break — sentence end, then clause end — so a trimmed reply is never half a sentence.
- Deterministic — one character, always on one side and never the other.
Close it. Terminal
?, double quotes, markdown*. - Statistical — overlapping distributions. Shape it at generation, and only ever trim at a clean boundary. Truncating mid-sentence is a louder tell than the one it removes.
One statistical gap is left open deliberately. Measured over 20 exchanges, distinct reply openings: ELIZA 95%, the model 23%. ELIZA cycles pointers through 199 hand-written templates; a 0.6B model has one system prompt and a sampling distribution. It cannot be tuned away, and it points the wrong way — the repetitive, wordier host is the 2026 one, which is the opposite of what a player expects. It makes the test harder, not easier.
It deliberately does not mention ELIZA, DOCTOR or 1966, does not ask the model to sound old-fashioned, and gives it no examples to imitate. Telling it to impersonate ELIZA would make this a test of impersonation. The question is whether sixty years is legible when both sides do the same job under the same constraints.
v1.0 — both hosts are real, the deployment runs, and the blind test is a blind test.
| Machine, sound, keyboard, print, i18n, scoring | done |
| ELIZA | done — DOCTOR script, 1966, with rank, cycling and MEMORY |
| LLM | done — qwen3:0.6b via llama-server, same origin |
| Deployment | docker compose up --build |
ELIZA is the real thing: the DOCTOR script from the CACM appendix, embedded
verbatim and parsed at load time rather than transcribed into JavaScript, so
that fidelity is a property of the file rather than of someone's proofreading.
npm test reproduces the conversation printed in the 1966 paper, exchange for
exchange.
A machine with Docker, three cores and 4 GB of RAM. No GPU, no API key, no account anywhere. The model runs on the CPU, and at 10 characters per second the paper is a far tighter bottleneck than the processor.
git clone <this repo>
cd 33talk
docker compose up --build -dOpen http://<your-server>:16001. That is the whole deployment.
The first build downloads the model — 429 MB, a few minutes. Nothing is downloaded at boot, and nothing is downloaded on later builds unless you change the model.
In mainland China, or anywhere huggingface.co is unreachable, build
through a mirror:
docker compose build --build-arg HF_HOST=https://hf-mirror.com
docker compose up -dThere is none by default. Everything below is optional — copy
.env.example to .env only if you need one of them.
PORT |
16001 |
Published port. Change if 16001 is taken |
MODEL_REPO |
ggml-org/Qwen3-0.6B-GGUF |
HuggingFace repo |
MODEL_FILE |
Qwen3-0.6B-Q4_0.gguf |
File within it |
HF_HOST |
https://huggingface.co |
Set to https://hf-mirror.com behind the GFW |
Changing the model is a rebuild, not a restart:
docker compose build llm && docker compose up -d.
| Logs | docker compose logs -f |
| Stop | docker compose down |
Update after a git pull |
docker compose up --build -d |
| Rebuild the front end only | docker compose build web — seconds, the model layer is cached |
Port already in use. echo "PORT=16005" > .env and docker compose up -d.
The build stalls or resets while downloading the model. It resumes:
the download uses curl -C - inside a retry loop, and verifies the finished
size against what the server advertised. If it fails ten times, use the
mirror above.
llama-server exits immediately with cannot open shared object file.
Something changed the container's working directory. It must stay /app —
the loader finds the .so files relative to it, and the error names a
missing library rather than the real cause.
Linux: host.docker.internal does not resolve. Only relevant if you
point the app at an inference server running on the host instead of the one
in Compose. compose.yml already carries the extra_hosts entry that fixes
it; without it you get a connection failure with no useful error.
browser --> nginx :16001 --+-- / --> web/ (static files)
+-- /v1/... --> llama-server :16002
Two containers. Same origin, so the browser never makes a cross-origin
request and CORS never enters the picture. There is no backend code — nginx
is configuration, not a program. llama-server is not published to the
host: the only way in is through the /v1 proxy.
| image | size | |
|---|---|---|
33talk-web |
48 MB | nginx + web/ |
33talk-llm |
555 MB | llama-server (39 MB) + model (429 MB) on ubuntu:24.04 |
The model is baked into the image. That reverses the usual advice, and it is affordable only because the model is tiny. The blind test is worthless if the LLM half is missing, so a deploy that boots and then spends minutes downloading is the worse trade.
The upstream llama.cpp:server image is 845 MB, of which the binaries are
39 MB and the rest is a build environment. A multi-stage build takes the 39
MB onto a plain ubuntu:24.04 and drops the image from 1.27 GB to 555 MB.
It has to be ubuntu:24.04 specifically — debian:12-slim is the same size
but two glibc versions older, and fails at load with GLIBC_2.38 not found.
llama-server can also serve a static directory itself (--path, verified
against the binary; --cors-origins defaults to *), so a single-process
deployment is possible. The two-service split is kept because it leaves the
inference container replaceable.
- Not a Model 33 emulator.
Random832/ttyemu,hughpyle/ttyemu,progs-n-things/asr33emuand the PDP-8/E simulator already do that, better, with paper tape and serial backends. Feature parity with them is explicitly not a goal. - Not a terminal. There is no shell, no PTY, no SSH.
- Not a chat UI with a retro skin.
"AI on a teletype" is not novel — m15-ai/Paper-Tape-LLM does it, and
hughpyle has run an LLM on a real ASR-33. The 1966-versus-2026 blind test
is.
Apache-2.0. See LICENSE and NOTICE.
ELIZA's original source is CC0 / public domain — see docs/eliza.md.
"Teletype" is a trademark, which is why this project is called 33talk and not teletype-anything.